Prolonged exposure to elevated ozone levels poses significant health risk. Accurate forecasting of ozone concentration is crucial for enabling timely interventions to mitigate their decremental effects on public health and environmental well-being. Long short-term memory (LSTM) networks have garnered substantial attention in ozone prediction due to their proficiency in capturing long-term temporal dependencies. Nonetheless, these models often overlook short-term non-stationarity resulting from periodic ozone trends, which remains a critical limitation. Existing studies predominantly address short-term ozone forecasting (within hours), thereby constraining early warning systems and impeding effective control measures. To overcome these limitations, we propose a hybrid deep learning model, DLinear-LSTM, tailored for 24-hour ozone concentration forecasting. The model leverages the moving average kernel decomposition module from DLinear to segregate time series data into trend and seasonal components. Subsequently, stacked LSTM units independently model these components, effectively capturing both long-term temporal patterns and periodic fluctuations in ozone levels. Empirical evaluations reveal that DLinear-LSTM achieves superior forecasting accuracy with a mean absolute error (MAE) of 19.15 µg/m³, root mean square error (RMSE) of 25.87 µg/m³, and coefficient of determination (R²) of 0.715, outperforming state-of-the-art forecasting models. Additionally, ablation studies underscore the essential role of time series decomposition in enhancing the performance of air quality prediction models. In summary, the proposed DLinear-LSTM framework offers a robust data-driven approach for accurate air pollutant levels forecasting, aiding atmospheric pollution control and safeguarding public health.
Cang et al. (Wed,) studied this question.